ExcaScene benchmark data and models: paired camera-LiDAR excavator perception under sensor faults (v1.0.0)

Data, checkpoints, per-decision outputs and numerical results for ExcaScene, a paired simulation benchmark that evaluates excavator perception by the decisions it leads to (regret and unsafe selections) as well as by map error. Includes perception checkpoints (KinAfford dynamic and BEVFusion-Exca), the camera-health gate, holdout seeds, exact array hashes for regenerating the two prospective holdouts' latent fields, and rendered evaluation observations, per-decision outputs, bootstrap multiplicities, complete study tables, protocols and locks. Code: https://github.com/nvh2520/excascene (Apache-2.0). GOOSE-Ex images are not redistributed. CC BY 4.0 applies to original ExcaScene material. The synthetic-domain segmenter initialized from NVIDIA MiT-B0 retains the upstream NVIDIA Source Code License restriction, including non-commercial use; see the README and source notices. GOOSE-Ex-trained weights are excluded.